At present,most video frame interpolation technologies have achieved good results,but due to the lack of 4K video data sets,these methods are not ideal for ultra-high-definition video processing.In order to solve the above prob-lems,this paper creates an ultra-high-definition video dataset UHD4K120FPS.At the same time,for 4K video,this paper proposes a video frame interpolation model based on kernel estimation and optical flow estimation.Specifically,the two in-put frames are respectively input into the kernel estimation sub-network and the optical flow estimation sub-network to ex-tract the features of kernel estimation and optical flow estimation,and the extracted features are processed and input into the post-processing fusion sub-network,and warped by cubic convolution and multiple convolutions output the final result.In this paper,training and verification tests are carried out on different data sets.
关键词
视频插帧/深度学习/光流法/4K视频
Key words
video frame interpolation/deep learning/optical flow/4K video
维普
万方数据